Federal learning-based multivariate data user portrait implementation method

Abstract

The invention discloses a multivariate data user portrait implementation method based on federal learning, and relates to the technical field of federal portray.The method comprises the steps that hierarchical grouping is conducted according to a cross-institution capacity registry to form collaborative sub-clusters, and a collaborative plan and privacy budget distribution table is generated according to capacity distribution of the collaborative sub-clusters; a capability calling list is generated and issued through a collaborative plan, each participating mechanism performs privacy processing on local portrait capability information according to a privacy budget allocation table, generates a local contribution fragment, uploads the local contribution fragment to a collaborative sub-cluster, and uploads the local contribution fragment to a federal coordination server after fusion of the collaborative sub-cluster is completed; the federal coordination server generates a global portrait capability result after completing global fusion and issues the global portrait capability result; through a hierarchical grouping mechanism based on a cross-mechanism capability registry, mechanisms with similar computing power, network conditions and data quality are clustered to form a collaborative sub-cluster, and load imbalance and communication bottleneck caused by mechanism heterogeneity in a traditional scheme are effectively eliminated.

Description

Technical Field

[0001] This invention relates to the field of federated profiling technology, and in particular to a method for implementing multi-data user profiling based on federated learning. Background Technology

[0002] With the rapid development of the digital economy, user profiling technology has become a core support for business scenarios such as financial risk control, precision marketing, and personalized recommendations. Early user profiling primarily relied on data from single institutions, directly constructing tag systems through machine learning or rule engines, achieving high accuracy. However, increasingly stringent requirements for personal information and data security compliance have placed serious privacy risks and legal compliance obstacles on traditional centralized data aggregation models. To address the problem of cross-institutional data silos, federated learning, as a distributed privacy-preserving computing paradigm, has been introduced into the field of user profiling. Existing user profiling solutions based on federated learning mainly include three categories: horizontal federation, vertical federation, and federated transfer learning. Among these, vertical federation has received widespread attention due to its suitability for cross-institutional scenarios with minimal overlap in feature dimensions and significant overlap in users.

[0003] However, there are still areas for improvement in the existing methods. First, the existing solutions generally lack a hierarchical consideration of the computing power, network conditions, and data quality of the participating institutions, which leads to load imbalance, communication bottlenecks, and low-quality data dragging down the overall profiling effect when there are a large number of institutions. In addition, most solutions still take joint training of a unified model as the core path, which leads to convergence difficulties and a significant drop in profiling accuracy when there is a large degree of heterogeneity. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a multi-data user profiling method based on federated learning to solve the problem of a lack of hierarchical consideration of the computing power, network conditions, and data quality of participating institutions, and the fact that most solutions still take joint training of a unified model as the core path, which leads to a significant decrease in profiling accuracy.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for implementing multi-dataset user profiling based on federated learning, which includes: The federal coordination server establishes a profile task directory and distributes capability description templates. Each participating agency registers its local profile capability information and uploads a capability description form, generating a cross-agency capability registry. Based on the cross-agency capability registry, hierarchical grouping is performed to form collaborative sub-clusters, and a collaborative plan and privacy budget allocation table are generated according to the capability distribution of the collaborative sub-clusters. The collaborative plan generates and distributes a list of capabilities to be invoked. Each participating organization performs privacy processing on its local profile capability information according to the privacy budget allocation table, generates a local contribution fragment, and uploads it to the collaborative sub-cluster. After the collaborative sub-cluster completes the fusion, it is uploaded to the federated coordination server. After the federated coordination server completes the global fusion, it generates and distributes the global profile capability result. Each participating organization completes local adaptation by combining the local contribution fragment. The global profile capability result is a structured result set that is indexed by the target user identifier and contains capability identifiers and corresponding capability values. The federal coordination server generates task-level user profiles based on the global profiling capabilities and returns them to the business functions. It also receives usage feedback records from the business functions to form an application result set.

[0007] As a preferred embodiment of the multi-database user profiling method based on federated learning described in this invention, the participating institutions register local profiling capability information and upload capability description tables to generate a cross-institutional capability registry, specifically: After receiving the capability description template, each participating institution will map its local multi-source data into multiple local profile capability information centers. For each local profile capability information center, calculate the number of users covered, data freshness, and historical accuracy, and generate capability description entries; The local federal agent adds an organization identifier to all capability description entries, encrypts and signs them, generates a capability description table, and uploads it to the federal coordination server through a secure channel. After receiving the capability descriptions from all participating agencies, the federal coordination server decrypts all capability descriptions, organizes and stores them according to uniform fields, forming a cross-agency capability registry.

[0008] As a preferred embodiment of the multi-database user profiling method based on federated learning described in this invention, the step of forming collaborative sub-clusters by hierarchical grouping according to the cross-institutional capability registry specifically includes: The federal coordination server extracts the capability type vector, privacy sensitivity level, computing power parameters, and network latency of each agency to obtain multiple collaborative sub-clusters; Assign a unique sub-cluster coordinating node to each collaborative sub-cluster, save the partitioning results of all collaborative sub-clusters as a collaborative sub-cluster partitioning table, and update it to the cross-organization capability registry.

[0009] As a preferred embodiment of the multi-dataset user profiling method based on federated learning described in this invention, the generation of the collaborative plan and privacy budget allocation table specifically includes: The federal coordination server generates a capability list for each collaborative sub-cluster based on the capability types required for each portrait task in the portrait task catalog and the capability coverage of the collaborative sub-cluster. Based on the privacy sensitivity level of each capability in the capability list and the real-time requirements of the profiling task, differential privacy noise intensity is assigned to the current capability to generate a task-level privacy budget allocation table. Package the capability list for profiling tasks and the corresponding task-level privacy budget allocation table into a collaborative plan.

[0010] As a preferred embodiment of the multi-dataset user profiling method based on federated learning described in this invention, wherein: each participating institution generates a local contribution fragment after performing privacy processing on its local profiling capability information according to the privacy budget allocation table, specifically: After receiving the capability call list, each participating organization extracts the local raw data of the specified user and the specified capability from the capability call list; Privacy processing is performed on the extracted local raw data according to the privacy budget allocation table. Laplace noise is added to continuous capabilities, exponential randomization is performed on categorical capabilities, and local hash perturbation is performed on sparse high-dimensional capabilities. The perturbed capability values ​​are then encapsulated and encrypted as local contribution fragments and uploaded to their respective collaborative sub-clusters.

[0011] As a preferred embodiment of the multi-dataset user profiling method based on federated learning described in this invention, the step of uploading the data to the federated coordination server after the collaborative sub-clusters have completed fusion specifically includes: The collaborative sub-cluster receives all local contribution fragments from its own cluster, performs a merging operation on the local contribution fragments through the collaborative plan, and obtains the preliminary fusion result of the sub-cluster. Noise is injected again into the initial fusion results of the sub-clusters to generate fused sub-cluster fragments, which are then uploaded to the federated coordination server.

[0012] As a preferred embodiment of the multi-dataset user profiling method based on federated learning described in this invention, the federated coordination server generates a global profiling capability result after completing global fusion, specifically as follows: The federal coordination server receives all sub-cluster fusion fragments and performs weighted aggregation on all sub-cluster fusion fragments through a coordination plan to obtain the aggregation result. The aggregation results are truncated by a threshold, normalized, and timestamped to generate a global profile capability result. The global profile capability result is then encrypted and distributed to each participating institution.

[0013] As a preferred embodiment of the multi-data user profiling method based on federated learning described in this invention, the local adaptation of each participating institution by combining local contribution fragments means that each participating institution receives the global profiling capability result, performs a weighted combination of the global profiling capability result and its own unnoised local contribution fragments, and then caches them.

[0014] As a preferred embodiment of the multi-dataset user profiling method based on federated learning described in this invention, the federated coordination server generates task-level user profiles based on global profiling capability results, specifically: The federal coordination server receives business function call requests in real time, and selects all capability tags and values ​​of the target user from the global profile capability results according to the combination rules of the profile task directory to form a capability subset. Priority arbitration is performed on conflicting labels in the capability subset, and task-level user profiles are generated after exponential decay of lagging capability labels.

[0015] As a preferred embodiment of the multi-dataset user profiling method based on federated learning described in this invention, wherein: the usage feedback records received from the service function are used to form an application result set, specifically: After completing business decisions using task-level user profiles, business functions generate usage feedback records and send these records back to the federated coordination server via a secure channel. The federal coordination server uses the profile task identifier and the target user identifier as keys to associate the usage feedback records with the corresponding task-level user profiles one by one; The associated records, along with the version number of the global profiling capabilities used at the time, are aggregated to form an application result set.

[0016] The beneficial effects of this invention are as follows: By using a hierarchical grouping mechanism based on a cross-institutional capability registry, institutions with similar computing power, network conditions, and data quality are clustered into collaborative sub-clusters, effectively eliminating the load imbalance and communication bottlenecks caused by institutional heterogeneity in traditional solutions, while avoiding low-quality data dragging down the overall effect; by abandoning the traditional unified model training path of multi-round gradient interaction and adopting a three-level fusion mechanism of "local contribution fragment → fusion within sub-clusters → global fusion → local adaptation", it can still converge quickly and output high-precision profiles even when the feature space is highly heterogeneous, solving the convergence difficulties and profile accuracy degradation problems of existing federated learning solutions in scenarios with many participating institutions and severely heterogeneous data. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for implementing multi-dataset user profiling based on federated learning.

[0019] Figure 2A flowchart for forming collaborative sub-clusters through hierarchical grouping.

[0020] Figure 3 This is a schematic diagram of the privacy processing and data fusion process.

[0021] Figure 4 A schematic diagram illustrating the generation and feedback of task-level user profiles. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for implementing multi-dataset user profiling based on federated learning, including the following steps: S1: The federal coordination server establishes a profile task directory and distributes capability description templates. Each participating agency registers its local profile capability information and uploads a capability description table, generating a cross-agency capability registry. S1.1: When the Federal Coordination Server begins creating the profiling task catalog, it lists all profiling task objectives one by one, specifying for each objective a task identifier (unique string), target user scope, required profiling capability type, task response time requirements, and a description of the task output structure. After compiling the profiling task catalog, the capability description template, used to describe profiling capabilities, is distributed to all participating organizations, enabling them to describe their local profiling capability information according to a unified field specification. The fields in the capability description template include a capability identifier field, a dependent data type field, an output tag meaning field, an update cycle field, a supported profiling task type field, and a privacy sensitivity level field.

[0026] It should also be noted that the profile task objective refers to the type of user profile that the Federal Coordination Server needs to build in various business scenarios in the management console and the application purpose. For example, the objective of the "Real-time Risk Control Profile Task" is to identify high-risk users, the objective of the "Near Real-time Recommendation Profile Task" is to generate interest preference tags, and the objective of the "Marketing Circle Selection Profile Task" is to segment high-value user groups, etc. The target user scope refers to the user set filtering criteria, i.e., a subset of user identifiers, such as "all active users" or "users with transaction records in the past 30 days." The required profiling capability type refers to the specific set of capability identifiers, which are pre-specified for the current profiling task in the profiling task directory. This set clarifies the mandatory or optional profiling capabilities for the current task; for example, a real-time risk control profiling task requires "spending capacity," "activity level," and "risk scoring capability." The task response time requirement refers to the maximum allowable latency; for example, "≤500 milliseconds" for a real-time risk control profiling task and "≤5 seconds" for a near real-time recommendation profiling task. The task output structure specification refers to the output format standard, such as "a JSON object containing user identifiers, an array of capability tags (each tag includes capability identifier, value, and confidence level), and a generation timestamp." These settings are configured by business operations personnel in the management console of the federated coordination server according to the actual business scenario, ensuring that each profiling task objective has a clear and executable definition.

[0027] Capability description templates are pre-generated and maintained by the federal coordination server to unify the description format of local profiling capability information for all participating institutions. During the initialization phase, the federal coordination server is configured by the administrator in the management console, defining and generating capability description templates based on the common capability dimensions required for all profiling tasks in the profiling task directory. Essentially, a capability description template is a standardized structured table, ensuring that the capability description tables uploaded by each institution have a consistent structure, facilitating subsequent aggregation and querying of cross-institutional capability registry entries. The specific field specifications in the capability description template are as follows: Capability Identifier field, used to fill in a unique string identifier; Dependent Data Type field, used to fill in the specific data source on which the current capability is based, such as "transaction flow" or "browsing logs"; Output Tag Meaning field, used to fill in the specific meaning of the final output tag or score of the current capability, such as "monthly total consumption score, range 0-100"; Update Cycle field, used to fill in the time interval for refreshing the current capability data, such as "daily" or "real-time"; Supported Profiling Task Type field, used to fill in the list of task identifiers that the current capability can serve; Privacy Sensitivity Level field, used to fill in a numerical value or level, such as "level 1-10, 10 is the most sensitive".

[0028] After receiving the capability description template, each participating institution will systematically review its local multi-source data based on the fields in the template, and map different information types in the local multi-source data to multiple local profile capability information centers. Specifically, the participating institutions will identify different information types in the local multi-source data, such as transaction records belonging to the consumption information type, access logs belonging to the behavioral preference information type, device information belonging to the terminal attribute information type, and location trajectory belonging to the geographic activity information type, and map each information type separately to a local profile capability information center. For example, transaction records are mapped to the consumption capability center, access logs are mapped to the interest preference capability center, device information is mapped to the device activity capability center, and location trajectory is mapped to the regional activity capability center. Each local profiling capability information center fills in the fixed fields specified in the capability description template according to the requirements. These requirements include: filling in the capability identifier field to record a unique identifier string; filling in the dependent data type field to record the specific data source on which the current local profiling capability information center is based; filling in the output tag meaning field to record the specific explanation of the tags or scores finally generated by the current local profiling capability information center; filling in the update cycle field to record the data refresh frequency of the current local profiling capability information center; filling in the supported profiling task type field to record the list of profiling task identifiers that the current local profiling capability information center can serve; and filling in the privacy sensitivity level field to record the sensitivity value or level of the current local profiling capability information center. Because all local profiling capability information centers strictly follow the fixed fields specified in the unified capability description template, each local profiling capability information center obtains standardized description information after completion, enabling it to have complete information for subsequent registration.

[0029] It should also be noted that local multi-source data refers to heterogeneous data resources held internally by participating institutions, including but not limited to structured or semi-structured data such as transaction logs, access logs, device information, location tracking, social interaction records, consumption records, and credit history. This data originates from different business systems and typically has different field formats, update frequencies, and privacy sensitivities, serving as the foundational information source for building user profile tags.

[0030] S1.2: After establishing their local profiling capability information centers, each participating institution will confirm the user coverage for each center: This is done by counting the number of users in the data types that the local profiling capability information center depends on; each participating institution will calculate the data freshness of the current local profiling capability information center by calculating the time difference between the current time and the latest timestamp of the data types it depends on, using the following formula: ; in, This value represents the data freshness of the local profiling capability information center, ranging from 0 to 1. A value closer to 1 indicates fresher data, while a value closer to 0 indicates older data. This range is determined based on the properties of an exponential function. This represents the time difference between the current moment and the latest timestamp of the data type on which the local profiling capability information center relies. This indicates the pre-set freshness decay constant for the current dependent data type, in hours, for example, 360 hours. It should also be noted that the freshness decay constant is set based on the natural update frequency of the data type and the sensitivity of the business scenario to timeliness.

[0031] By comparing the tag output of the local profile capability information center in historical business with the actual business performance, the historical accuracy of the local profile capability information center is obtained. The number of users covered, data freshness, and historical accuracy are combined with the basic fields of the local profile capability information center (capability identifier, dependent data type, meaning of output tags, update cycle, supported profile task types, and privacy sensitivity level) to form a complete capability description item.

[0032] Each participating agency's local federal agent collects all capability description entries for their agency and adds an agency identifier field before each entry, enabling the federal coordination server to identify the source of the capability description entry. The local federal agent performs an encrypted signature operation on the complete file containing all capability description entries, ensuring the integrity and immutability of the capability description table content during the upload process. After the encryption and signing are completed, the capability description table is uploaded to the federal coordination server through a secure channel.

[0033] It should also be noted that when generating the capability description table, each participating institution, in addition to the capability description entries, also attaches its own metadata, including computing power parameters (based on the standardized score of local computing resources), network latency (average round-trip latency with the federal coordination server), and a capability type vector formed by summarizing the task types supported by all local profiling capability information centers. When the federal coordination server summarizes the cross-institutional capability registry, it stores the above-mentioned institutional metadata together with the privacy sensitivity level field to form a complete institutional characteristic description for subsequent hierarchical grouping.

[0034] After receiving the capability descriptions from all participating organizations, the federal coordination server organizes them uniformly according to fields such as organization identifier, capability identifier, dependent data type, output tag meaning, update cycle, supported profiling task types, privacy sensitivity level, number of users covered, data freshness, and historical accuracy, and summarizes them into a cross-organization capability registry.

[0035] S2: Based on the cross-agency capability registry, hierarchical grouping is performed to form collaborative sub-clusters, and a collaborative plan and privacy budget allocation table are generated based on the capability distribution of the collaborative sub-clusters; S2.1: After receiving the cross-agency capability registry, the federal coordination server sequentially reads the capability type vector field, privacy sensitivity level field, computing power parameter field, and network latency field of each participating agency in the cross-agency capability registry, and uses the above four types of fields as the basis for grouping to form a set of input information for subsequent division, so that subsequent hierarchical grouping can be carried out based on a unified data foundation.

[0036] After obtaining the input information set, the capability type vector fields of each participating organization are compared dimension by dimension when matching them. Each capability identifier in the capability type vector field is considered an independent dimension; for example, consumption capability identifier is the first dimension, content preference capability identifier is the second dimension, and so on. During the comparison, the federated coordination server counts the number of identical capability identifiers in the capability type vector fields of two participating organizations and uses this number as the similarity score between the two organizations in the capability type vector field. For example, the similarity score equals the number of shared capability identifiers between the two organizations. A higher similarity score indicates that the two participating organizations can undertake more common capabilities in the same type of profiling task. The federated coordination server prioritizes grouping participating organizations with high similarity scores in the capability type vector fields into the same initial collaborative sub-cluster.

[0037] After completing the initial classification of the capability type vector field, the federated coordination server performs consistency screening on the privacy sensitivity level field, computing power parameter field, and network latency field within each initial collaborative sub-cluster. The consistency rule for the privacy sensitivity level field is: when the difference between the privacy sensitivity level fields of two participating institutions is less than a preset privacy threshold, the privacy sensitivity level fields of the two participating institutions are considered to match. The consistency rule for the computing power parameter field is: when the computing power parameter fields of two participating institutions are within the same order of magnitude range, the computing power parameter fields of the two participating institutions are considered to match. The same order of magnitude range means that the computing power parameter values ​​are within the same logarithmic level interval, such as 0-10, 10-100, 100-1000, etc. The matching rule for the network latency field is: when the network latency fields of two participating institutions are within the same time period range, the network latency fields of the two participating institutions are considered to match. The same time period range means that the network latency values ​​are within the same interval, such as 0-50 milliseconds, 50-200 milliseconds, 200-1000 milliseconds, etc. Based on the matching results of three types of fields—privacy sensitivity level, computing power parameter, and network latency—the federal coordination server retains participating organizations with a high degree of matching in the same initial coordination sub-cluster, thus forming a set of members that meet privacy protection requirements, have comparable computing power, and are compatible with communication conditions.

[0038] It should also be noted that the preset privacy threshold is generated by the federal coordination server from all privacy-sensitive level fields in the cross-agency capability registry. The value is the sum of the average of the set of privacy-sensitive level fields and the standard deviation of the set. For example, the value typically falls within the upper third of the overall range of privacy-sensitive level fields. This is set to objectively reflect the upper limit of the distribution of privacy-sensitive level fields across participating agencies. This calculation method was chosen as the preset privacy threshold because, assuming the privacy-sensitive level fields approximately follow a normal distribution, this method can objectively cover the upper limit of agency distribution, avoiding bias caused by subjective settings and effectively highlighting agencies with high protection needs at the tail end of the distribution. A high degree of matching means that if two participating agencies meet the matching requirements in at least two of the three fields—privacy-sensitive level field, computing power parameter field, and network latency field—and the difference between the privacy-sensitive level fields of the two participating agencies does not exceed the preset privacy threshold, then the requirement of a high degree of matching is met.

[0039] When the federal coordination server confirms the consistency of privacy protection among members of the collaborative sub-cluster, if the privacy sensitivity level field of a participating organization is higher than the preset privacy threshold, it must be retained in its respective collaborative sub-cluster. At the same time, it requires that the members of the same collaborative sub-cluster of the participating organization with a privacy sensitivity level field not lower than the current participating organization's privacy sensitivity level field, so as to ensure that the privacy protection needs of the participating organizations in the collaborative sub-cluster can be met.

[0040] After the members of the collaborative sub-clusters are determined, all participating institutions in each collaborative sub-cluster are sorted according to the computing power parameter field and the network latency field. Participating institutions with high computing power parameter field and low network latency field are selected as candidate coordination nodes. Specifically, the computing power parameter field is defined as follows: the higher the value, the stronger the computing power. The value is obtained by standardizing the calculation of each participating institution's local computing resources. For example, it is normalized to a range of 0-100 by taking into account factors such as the number of CPU cores, memory size, and GPU availability. The unit is points. A value above 80 points is considered high computing power (determined based on standard server load assessment. In cloud computing scheduling, a value above 80 points usually represents a high-performance node that can reliably undertake coordination tasks). The network latency field is defined as follows: the lower the value, the lower the latency. The value is expressed as the average round-trip latency in milliseconds. The unit is milliseconds. For example, a value below 50 milliseconds is considered low latency (determined based on real-time collaborative communication requirements. In low-latency scenarios such as financial risk control and recommendation, a value below 50 milliseconds is usually considered a good network condition to ensure efficient communication between coordination nodes). The federated coordination server selects the most suitable participating institution (the one with the highest value in the computing power parameter field and the lowest value in the network latency field) from the candidate list as the coordinating institution, and marks the selected participating institution as a sub-cluster coordination node. The federated coordination server saves the member list of each collaborative sub-cluster, the sub-cluster coordination node identifier, and the corresponding capability type vector field, privacy sensitivity level field, computing power parameter field, and network latency field as a collaborative sub-cluster partitioning table, and synchronously records the collaborative sub-cluster partitioning table to the cross-institutional capability registry, so that the cross-institutional capability registry becomes the basis for subsequent collaborative operations.

[0041] S2.2: After the collaborative sub-cluster partitioning table is generated, a basic strategy for the collaborative plan is generated based on the capability type vector field, privacy sensitivity level field, computing power parameter field, and network latency field recorded in the collaborative sub-cluster partitioning table. This strategy includes sub-cluster fusion rules and inter-sub-cluster weighting schemes. The sub-cluster fusion rules determine the processing method used when fusion occurs after receiving local contribution fragments within the collaborative sub-cluster. The processing method can be determined based on the degree of difference in the capability type vector field set, the protection requirements of the privacy sensitivity level field set, and the processing capacity range of the computing power parameter field set. Specifically, the processing methods are as follows: when the overlap ratio of capability identifiers supported by participating institutions within the sub-cluster exceeds 80% and the average privacy sensitivity level is below level 5, a weighted average method is used for fusion; when the overlap ratio of capability identifiers supported by participating institutions within the sub-cluster is below 50% or the average privacy sensitivity level is above level 7, a majority voting method is used for fusion; when the average computing power parameter score within the sub-cluster is above 70, a secure multi-party summation method is used for fusion.

[0042] It should also be noted that the score of over 80% is based on the impact of feature intersection rate on aggregation accuracy in vertical federated learning, while the score of less than 50% is based on benchmark tests of privacy computing frameworks (such as FATE) with lower privacy leakage risk in low intersection scenarios, indicating that the overlap is insufficient for reliable fusion and the protection mode needs to be switched; the score of over 70% is based on the computing power requirements of open source federated platforms for coordinating nodes, indicating that it is sufficient to handle complex fusion computations.

[0043] The weighting scheme among sub-clusters is used to determine the relative weights of multiple collaborative sub-clusters during global fusion. The weighting calculation formula is as follows: ; in, Indicates the first The weights of each collaborative sub-cluster Indicates the first The number of members in each collaborative sub-cluster is chosen because it reflects the data scale and representativeness of the sub-cluster. More members mean a larger user sample and a larger amount of data covered, and the contributed fragments should have higher credibility and statistical significance during global fusion. Indicates the first The average historical accuracy of the collaborative sub-clusters reflects the quality of historical predictions of the capabilities of participating institutions within the sub-cluster. A higher accuracy indicates stronger reliability of the contribution segments provided by the sub-cluster. Indicates the total number of collaborative sub-clusters. Indicates the first The number of participating organizations included in each collaborative sub-cluster. Indicates the first The arithmetic mean of the historical accuracy rates of all participating institutions within a collaborative sub-cluster.

[0044] It should also be noted that the product of the number of members and the average historical accuracy is used as the numerator, so that the weights take into account both "quantity" (scale) and "quality" (accuracy). This ensures that sub-clusters with rich and high-quality data dominate the global fusion results, which directly reflects the core logic of the weighting scheme among sub-clusters: by quantifying the comprehensive indicators of scale and quality, the global fusion is tilted towards high-quality contributions, thereby maximizing the profiling effect in scenarios involving multiple heterogeneous institutions.

[0045] The formula calculates that collaborative sub-clusters with a larger number of members and higher historical accuracy will receive greater weight. After completing the fusion rules within sub-clusters and the weighting scheme between sub-clusters, these rules will be used as the basis for generating the collaborative plan.

[0046] Given a catalog of profiling tasks and a table of collaborative sub-clusters, the federal coordination server examines each required capability type field for a profiling task in the catalog and compares it with the registered capability type fields within the collaborative sub-clusters to confirm whether the collaborative sub-clusters have the capability coverage to complete the profiling tasks. Capability coverage refers to the set of capability type fields registered by all participating institutions within the current collaborative sub-cluster that fully or partially includes the required capability type fields registered in the profiling task catalog for the current profiling task. This is obtained by the federal coordination server using the set of required capability type fields for the current profiling task (e.g., "consumption capacity," "interest preference capacity," "risk scoring capacity") as a standard set and comparing it with the capability type fields of all participating institutions within the collaborative sub-cluster. The server then calculates the proportion or number of capability types in the standard set that are supported by at least one participating institution within the collaborative sub-cluster, forming a coverage assessment result. Based on the comparison results of the coverage of capabilities to complete the portrait task, a capability list containing all task-related capability types is formed for each collaborative sub-cluster. Specifically: if the coverage ratio of the collaborative sub-cluster reaches 100%, the current collaborative sub-cluster is marked as having full capability coverage, the formed capability list contains all task-related capability types, and portrait tasks are assigned with priority. If the coverage ratio is less than 50%, it is marked as not having capability, an empty capability list is formed or no list is generated, and the current collaborative sub-cluster does not participate in the collection of contribution fragments for the current portrait task. If the coverage ratio is between 50% and 99%, the current collaborative sub-cluster is marked as having partial capability coverage, the formed capability list only contains the covered task-related capability types, and the missing capability types are recorded for the federated coordination server to supplement from other collaborative sub-clusters.

[0047] The privacy sensitivity level field of each capability in the capability list is read, and combined with the real-time requirement field of the corresponding portrait task in the portrait task directory, a differential privacy noise intensity is assigned to each capability using the existing differential privacy noise control method. Specifically, when the privacy sensitivity level field value is in the range of 1-10, a noise intensity of 0.5 is assigned for values ​​≥7, 1.0 for values ​​5-6, and 2.0 for values ​​≤4. When the real-time requirement field is "real-time" (response time ≤ 500 milliseconds), the noise intensity value is multiplied by 0.8; when the real-time requirement field is "near real-time" (response time ≤ 5 seconds), the original noise intensity value is maintained; and when the real-time requirement field is "batch" (response time > 5 seconds), the noise intensity value is multiplied by 1.2, thus obtaining the differential privacy noise intensity result. The federated coordination server compiles the differential privacy noise intensity results for each capability into a task-level privacy budget allocation table.

[0048] It should also be noted that the base noise intensity corresponding to the privacy sensitivity level classification is determined based on the inverse relationship between privacy budget and protection strength in differential privacy theory; the smaller the privacy budget, the stronger the privacy protection. The multiplication coefficient corresponding to the real-time requirement is determined based on the privacy-utility trade-off and task time constraints.

[0049] After obtaining the capability list and the task-level privacy budget allocation table, the capability list of the profiling task, the fusion rules within the sub-cluster, the weight scheme between collaborative sub-clusters, and the task-level privacy budget allocation table are integrated to form a complete collaborative plan.

[0050] Preferably, compared to traditional methods, this invention introduces a hierarchical grouping mechanism based on capability type vector fields, privacy sensitivity level fields, computing power parameter fields, and network latency fields. This enables collaborative sub-clusters to achieve refined matching among participating institutions with significant differences in privacy protection requirements, computing power capacity, and communication conditions. Simultaneously, the sub-cluster fusion rules generated from the field set and the inter-sub-cluster weighting scheme give the collaborative plan interpretability and dynamic adaptability, thereby improving the stability, efficiency, and privacy security of cross-institutional collaborative profiling tasks.

[0051] S3: A capability call list is generated and distributed through the collaborative plan. Each participating institution performs privacy processing on the local profile capability information according to the privacy budget allocation table, generates a local contribution fragment, and uploads it to the collaborative sub-cluster. After the collaborative sub-cluster completes the fusion, it is uploaded to the federated coordination server. After the federated coordination server completes the global fusion, it generates the global profile capability result and distributes it. Each participating institution completes local adaptation by combining the local contribution fragment. S3.1: Upon receiving the capability call list (i.e., capability list) generated by the collaboration plan, each participating institution sequentially reads the unified user identifier field and unified capability identifier field recorded in the capability call list locally to clarify the source of the local raw data to be processed. Each participating institution searches for the local raw data content corresponding to the current unified user identifier in its local multi-source data set based on the specified unified user identifier field. This search is performed using a pre-established cross-institutional unified user identifier to local user identifier mapping table (generated by the federated coordination server during initialization using a privacy-preserving ID alignment method to ensure that local user identifiers from different institutions are mapped to the same unified user identifier without data leaving the domain). Finally, based on the unified capability identifier field, the participating institution locates the local attribute field associated with the current unified capability, forming a local raw data set for privacy processing.Each participating institution selects a privacy processing method based on the privacy budget field and the privacy sensitivity level field corresponding to its current capability, as recorded in the task-level privacy budget allocation table: When the capability corresponding to the local raw data set is a continuous numerical type (such as consumption score, activity score, etc., which are real numbers), privacy protection is achieved by adding noise values ​​generated by the existing differential privacy mechanism to the continuous numerical value; when the capability corresponding to the local raw data set is a finite category (such as high / medium / low risk levels, interest category tags, etc., which are enumerated values), the category selection probability is perturbed through an index mechanism, so that the output category is randomly selected according to the index weight. Specifically: each participating institution calculates the original utility score of each candidate category based on the local raw data. The original utility score reflects the degree of matching between the category and the data; the index weight is calculated for each candidate category, that is, the original utility score is multiplied by half of the privacy budget and the natural exponent is taken; the index weights of all candidate categories are normalized to obtain the perturbed selection probability; a category is randomly sampled as the output result according to the perturbed selection probability, so that the category with high original utility has a higher probability of being selected, but controllable randomness is introduced to prevent direct leakage of the original data, thereby While preserving the practicality of category output, differential privacy protection requirements are met. When the capabilities corresponding to the local original dataset are sparse high-dimensional capabilities (such as one-hot encoded interest vectors, behavior sequence embeddings, etc., vectors with dimensions exceeding 1000 and sparsity higher than 90%), the existing locality-sensitive hash perturbation method is used to bucket the sparse high-dimensional features, mapping similar features to the same bucket and adding random offsets. Specifically, multiple random hyperplanes are selected as hash functions, and the sparse high-dimensional feature vectors are projected onto multiple random hyperplanes to obtain binary hash signatures; the hash signatures are then... Sparse high-dimensional feature vectors are mapped to the same hash bucket to achieve clustering of similar features. Within each hash bucket, a random offset vector is added to the sparse high-dimensional feature vector. This random offset vector is sampled from a Gaussian or uniform distribution, and its amplitude is controlled by the noise intensity allocated in the task-level privacy budget allocation table. The offset-added sparse high-dimensional feature vector is output as the perturbation result, ensuring that similar features still have a high probability of falling into nearby hash buckets to preserve some similarity. Simultaneously, the random offset ensures that the original vector is difficult to recover in reverse, thus providing differential privacy protection and maintaining feature availability in high-dimensional sparse scenarios. All perturbed capability values ​​are fully encapsulated into structured data fragments and encrypted using encryption keys held locally by participating institutions, forming local contribution fragments. After encryption and encapsulation, each participating institution uploads its local contribution fragment to its respective collaborative sub-cluster via a secure communication link.

[0052] S3.2: After receiving the local contribution fragments uploaded by all members, the collaborative sub-cluster will perform processing operations on the local contribution fragments according to the sub-cluster fusion rules recorded in the collaboration plan. The content of the processing operations depends on the configuration of the capability type vector field set in the sub-cluster fusion rules. For example, when the fusion rule requires aggregation based on similar capability tags, the collaborative sub-cluster will uniformly merge the perturbed tags belonging to the same capability identifier field; when the fusion rule requires sorting or filtering tags from different sources, the collaborative sub-cluster will process them according to the content of the processing operations during the fusion process. After the processing operations are completed, the preliminary fusion result of the sub-cluster is formed. To further ensure cross-organizational privacy protection requirements, the collaborative sub-cluster will add noise to the preliminary fusion result of the sub-cluster based on the remaining privacy budget recorded in the task-level privacy budget allocation table, so that the preliminary fusion result of the sub-cluster has an additional layer of privacy protection before leaving the sub-cluster. The collaborative sub-cluster uploads the sub-cluster fusion fragment with added noise to the federated coordination server for constructing the global profile capability result.

[0053] S3.3: After receiving the sub-cluster fusion fragments uploaded by all collaborative sub-clusters, the federated coordination server reads the inter-sub-cluster weight scheme in the coordination plan and assigns a corresponding weight value to each sub-cluster fusion fragment, applying the same capability identifier... Perturbation values ​​within the same cooperative sub-cluster The final value of the computing capability identifier in the global profile capability result is calculated using the following formula: ; in, The capability identifier is The final value of the ability tag in the global profile ability result (i.e., the global aggregate value). Indicates the total number of collaborative sub-clusters. Indicates the sequence number of the collaborative sub-cluster. The weight scheme among sub-clusters in the collaborative plan is represented as the first... The weight value assigned to each collaborative sub-cluster depends on how many collaborative sub-clusters the current profiling task involves, with a total weight value of 1. Indicates the first In the sub-cluster fusion fragments uploaded by each collaborative sub-cluster, the capability identifier is... The perturbed value carried by the capability label.

[0054] After calculating all capability identifiers, the global aggregation result is obtained.

[0055] After obtaining the global aggregation result, the task-level label availability requirement field recorded in the collaboration plan is read. The label availability requirement field includes constraints such as the minimum label occurrence frequency requirement (no less than 50% of the total number of sub-clusters, set to ensure that capability labels appear in a sufficient number of collaboration sub-clusters to have statistical reliability), the minimum label weight requirement (set to 0.1, which is set to ensure that low-weight labels are not retained to avoid noise dominating the results), and the minimum label stability requirement (set to ensure that the standard deviation of label values ​​output across sub-clusters does not exceed 20% of the mean, which is set to ensure that labels do not fluctuate much between different sub-clusters to ensure consistency). The federated coordination server evaluates each capability tag in the global aggregation result, considering the weighting scheme between sub-clusters and the configuration of capability tags in the task-level privacy budget allocation table. It assesses the tag's frequency of occurrence in different collaborative sub-clusters (at least 50% of the total number of collaborative sub-clusters is considered satisfactory), its weight in the global aggregation result (at least 0.3 is considered satisfactory), and its fluctuation in multiple rounds of profiling tasks (the ratio of standard deviation to mean is no more than 0.2 is considered satisfactory). Based on this, a tag contribution index is generated for each capability tag. The tag contribution index reflects the effective support of the current capability tag in the current profiling task. The federated coordination server marks the tag contribution index of capability tags that meet the minimum frequency of occurrence, minimum tag weight, and minimum tag stability requirements as 1, and those that do not meet any of these requirements as 0. Only capability tags with a tag contribution index of 1 are retained in the global aggregation result, ensuring that every capability tag in the global profiling capability result has sufficient task support value.

[0056] To ensure consistency in the scale of each capability label, the federated coordination server performs normalization on the truncated values ​​and adds a timestamp field to the global aggregation result after normalization to record the generation time of the global profiling capability result. The federated coordination server fully encrypts the generated global profiling capability result and distributes it to all participating institutions through a secure communication link, enabling participating institutions to improve their local profiling capability information based on the global profiling capability result.

[0057] S3.4: After receiving the global profiling capability results, each participating institution shall sequentially read each capability label in the global profiling capability results and search for the local reference data required for the corresponding capability label from the noisy local contribution fragments cached by the institution.

[0058] Each participating organization adapts according to the local adaptation rules in the collaboration plan. During local adaptation, the global weight field and local weight field recorded in the local adaptation rules are read sequentially, and the corresponding tag value of each capability tag in the global profile capability result is compared with the corresponding tag value in the noisy local contribution segment of the organization. For the same capability tag, the value in the global profile capability result and the value in the noisy local contribution segment are extracted. Each participating organization calculates the weighted value from the global profile capability result (multiplication of the global weight field value recorded in the local adaptation rules with the value of the current capability tag in the global profile capability result) and the weighted value from the noisy local contribution segment (multiplication of the local weight field value recorded in the local adaptation rules with the value of the same capability tag in the noisy local contribution segment cached by the organization) for each capability tag according to the ratio between the global weight field and the local weight field (the ratio is directly determined by the global weight field value and the local weight field value). The weighted value from the global profile capability result and the weighted value from the noisy local contribution segment are summed to form the local adaptation value of the current capability tag. The local adaptation value can simultaneously reflect the overall trend of the global profiling capability results across institutions as well as the authenticity of the un-noiseed local contribution fragments within the institution, enabling the final generated local enhanced profiling capability results to have both wide-area consistency and local accuracy.

[0059] Preferably, this invention achieves stable fusion results for multi-source contributions while maintaining privacy by performing differentiated privacy processing locally, implementing a fusion strategy based on capability features within collaborative sub-clusters, and performing weight-based global aggregation on the federated coordination server. Simultaneously, through label availability constraints and local adaptation mechanisms, a dynamic balance is achieved between cross-agency consistency and local accuracy in the global profiling results, thereby significantly improving the reliability and availability of the user profile generation process.

[0060] S4: The federal coordination server generates task-level user profiles based on the global profiling capabilities and returns them to the business functions. It also receives usage feedback records from the business functions to form an application result set. S4.1: The federated coordination server continuously listens for call requests from business functions. Upon receiving a call request containing a profile task identifier, a target user identifier set, and response time limits, it immediately reads the combination rules registered in the profile task directory for the profile task identifier. These combination rules are pre-configured when the profile task directory is created. The combination rules explicitly define the set of capability identifiers that the current profile task must include (such as core capabilities like risk scoring and spending power) and the set of optional capability identifiers (such as auxiliary capabilities like interest preferences and regional activity). Based on the combination rules, it extracts all capability tags and values ​​of the target user identifier set on the mandatory and optional capability identifier sets from the latest cached global profile capability results, forming the capability subset corresponding to the current call request.

[0061] Conflict detection is performed on all capability tags in the capability subset: When the same target user has multiple different values ​​under the same capability identifier, the priority order configured for each capability identifier in the profile task directory is used. The priority order is set according to business scenario requirements when the profile task directory is established. For example, the risk scoring capability is set to the highest priority in the real-time risk control profile task, and the interest preference capability is set to the highest priority in the near real-time recommendation profile task. The priority value ranges from 1 to 10, with higher values ​​indicating higher priority. The priority can be adjusted according to business changes. The capability tag value with the highest priority is retained, and the other conflicting values ​​are discarded. When the same capability identifier gives different values ​​in the sub-cluster fusion segments from different collaborative sub-cluster sources, the source with the highest sub-cluster weight scheme value assigned to the current collaborative sub-cluster in the collaborative plan is used. The value is retained, and values ​​from other sources are discarded.

[0062] S4.2: The federated coordination server continues to process the timeliness of capability tags in the capability subset: it reads the timestamp field carried by each capability tag in the global profile capability result, calculates the time difference between the current time and the timestamp field; if the time difference exceeds the maximum valid duration specified for the current capability identifier in the profile task directory (for example, for "Risk Scoring Capability" in "Real-time Risk Control Profile Task", the maximum valid duration can be set to 1 hour (3600 seconds), because risk scoring requires high real-time performance, and data exceeding 1 hour cannot reflect the current risk status), then the value of the current capability tag is reduced exponentially, and the reduced value is used to replace the original value. The expression is: ; in, This represents the attenuated value. Represents the original numerical value of the ability label. This represents the difference between the current time and the timestamp of the current capability tag. This indicates the preset half-life of the current capability identifier in the portrait task catalog, for example, 720 hours.

[0063] If the time difference does not exceed the maximum effective duration, the original value remains unchanged.

[0064] After completing the priority arbitration of conflict labels and the exponential decay of delayed capabilities, the federated coordination server organizes the processed capability subset (i.e., the capability labels and final value set after conflict value arbitration and time decay adjustment) into a complete task-level user profile according to the output structure specified in the profile task directory. The task-level user profile includes the target user identifier, profile task identifier, final capability labels and values, and generation timestamp. The task-level user profile is returned to the business function that initiated the call in real time through an encrypted channel.

[0065] After using task-level user profiles to complete risk control judgments, recommendation rankings, or marketing screenings, business functions associate the actual business results with the key capability tags in the task-level user profiles, generating a usage feedback record that includes only the profile task identifier, target user identifier, capability tags used, business success or failure markers, and whether the response time limit is met. The business functions then transmit this usage feedback record back to the federated coordination server via a secure channel.

[0066] It should also be noted that key capability tags refer to the core business decision-making capability tags marked for the current profile task in the profile task directory, such as risk scoring capability and fraud history capability in real-time risk control tasks, and interest preference capability and consumption capability in recommendation tasks.

[0067] After receiving all usage feedback records, the federal coordination server associates the usage feedback records with the corresponding task-level user profiles one-to-one using the profile task identifier and the target user identifier as keys. It also imports the global profile task identifier and capability subset version number used at that time into the application result set to form a complete record with a business verification tag.

[0068] Preferably, this invention enables task-level user profiles to output consistent and reliable results even in scenarios involving cross-organizational data conflicts and time delays by prioritizing conflicting labels and using exponential decay timeliness processing. At the same time, by using structured one-to-one associations of feedback records, a traceable set of application results is formed, avoiding the decrease in accuracy and auditing difficulties caused by version confusion or missing feedback in traditional federated profiles. This significantly improves the stability and practicality of profiles in business scenarios such as risk control and recommendation.

[0069] This embodiment also provides a computer device applicable to the implementation of a multi-dataset user profile method based on federated learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-dataset user profile implementation method based on federated learning as proposed in the above embodiment.

[0070] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0071] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the multi-data user profiling method based on federated learning proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0072] In summary, this invention effectively eliminates the load imbalance and communication bottlenecks caused by institutional heterogeneity in traditional solutions by using a hierarchical grouping mechanism based on a cross-institutional capability registry. This mechanism clusters institutions with similar computing power, network conditions, and data quality into collaborative sub-clusters and designates coordination nodes for these sub-clusters. It also avoids low-quality data dragging down the overall performance. Furthermore, by abandoning the traditional multi-round gradient interaction unified model training path and adopting a three-level fusion mechanism of "local contribution fragment → sub-cluster fusion → global fusion → local adaptation," this invention can still converge quickly and output high-precision profiles even in highly heterogeneous feature spaces. This solves the convergence difficulties and profile accuracy degradation problems of existing federated learning solutions in scenarios with numerous participating institutions and severely heterogeneous data.

[0073] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for implementing multi-dimensional data user portrait based on federated learning, characterized in that: include, The federal coordination server establishes a profile task directory and distributes capability description templates. Each participating agency registers its local profile capability information and uploads a capability description form, generating a cross-agency capability registry. Based on the cross-agency capability registry, hierarchical grouping is performed to form collaborative sub-clusters, and a collaborative plan and privacy budget allocation table are generated according to the capability distribution of the collaborative sub-clusters. The collaborative plan generates and distributes a list of capabilities to be invoked. Each participating organization performs privacy processing on its local profile capability information according to the privacy budget allocation table, generates a local contribution fragment, and uploads it to the collaborative sub-cluster. After the collaborative sub-cluster completes the fusion, it is uploaded to the federated coordination server. After the federated coordination server completes the global fusion, it generates and distributes the global profile capability result. Each participating organization completes local adaptation by combining the local contribution fragment. The global profile capability result is a structured result set that is indexed by the target user identifier and contains capability identifiers and corresponding capability values. The federal coordination server generates task-level user profiles based on the global profiling capabilities and returns them to the business functions. It also receives usage feedback records from the business functions to form an application result set.

2. The method of claim 1, wherein the method is implemented based on federated learning. Each participating institution registers its local profiling capabilities and uploads a capability description form, generating a cross-institutional capability registry, specifically as follows: After receiving the capability description template, each participating institution will map its local multi-source data into multiple local profile capability information centers. For each local profile capability information center, calculate the number of users covered, data freshness, and historical accuracy, and generate capability description entries; The local federal agent adds an organization identifier to all capability description entries, encrypts and signs them, generates a capability description table, and uploads it to the federal coordination server through a secure channel. After receiving the capability descriptions from all participating agencies, the federal coordination server decrypts all capability descriptions, organizes and stores them according to uniform fields, forming a cross-agency capability registry. 3.The method of claim 1, wherein: The hierarchical grouping based on the cross-agency capability registry to form collaborative sub-clusters specifically involves: The federal coordination server extracts the capability type vector, privacy sensitivity level, computing power parameters, and network latency of each agency to obtain multiple collaborative sub-clusters; Assign a unique sub-cluster coordinating node to each collaborative sub-cluster, save the partitioning results of all collaborative sub-clusters as a collaborative sub-cluster partitioning table, and update it to the cross-organization capability registry.

4. The method of claim 1, wherein: The generated collaboration plan and privacy budget allocation table are as follows: The federal coordination server generates a capability list for each collaborative sub-cluster based on the capability types required for each portrait task in the portrait task catalog and the capability coverage of the collaborative sub-cluster. Based on the privacy sensitivity level of each capability in the capability list and the real-time requirements of the profiling task, differential privacy noise intensity is assigned to the current capability to generate a task-level privacy budget allocation table. Package the capability list for profiling tasks and the corresponding task-level privacy budget allocation table into a collaborative plan.

5. The method for implementing multi-dataset user profiling based on federated learning as described in claim 1, characterized in that: Each participating organization generates a local contribution fragment after performing privacy processing on its local profile capability information according to the privacy budget allocation table. Specifically: After receiving the capability call list, each participating organization extracts the local raw data of the specified user and the specified capability from the capability call list; Privacy processing is performed on the extracted local raw data according to the privacy budget allocation table. Laplace noise is added to continuous capabilities, exponential randomization is performed on categorical capabilities, and local hash perturbation is performed on sparse high-dimensional capabilities. The perturbed capability values ​​are then encapsulated and encrypted as local contribution fragments and uploaded to their respective collaborative sub-clusters.

6. The method for implementing multi-dataset user profiling based on federated learning as described in claim 1, characterized in that: The process of uploading the data to the federated coordination server after the collaborative sub-clusters have completed fusion is as follows: The collaborative sub-cluster receives all local contribution fragments within its own cluster, and performs corresponding merging operations on the local contribution fragments through the collaborative plan to obtain the preliminary fusion result of the sub-cluster; Noise is injected again into the initial fusion results of the sub-clusters to generate fused sub-cluster fragments, which are then uploaded to the federated coordination server.

7. The method for implementing multi-dataset user profiling based on federated learning as described in claim 1, characterized in that: After completing global fusion, the federated coordination server generates a global profile capability result, specifically: The federal coordination server receives all sub-cluster fusion fragments and performs weighted aggregation on all sub-cluster fusion fragments through a coordination plan to obtain the aggregation result. The aggregation results are truncated by a threshold, normalized, and timestamped to generate a global profile capability result. The global profile capability result is then encrypted and distributed to each participating institution.

8. The method for implementing multi-dataset user profiling based on federated learning as described in claim 1, characterized in that: The phrase "each participating institution completes local adaptation by combining local contribution fragments" means that each participating institution receives the global profiling capability results, performs a weighted combination of the global profiling capability results and its own un-noiseed local contribution fragments, and then caches them.

9. The method for implementing multi-dataset user profiling based on federated learning as described in claim 1, characterized in that: The federated coordination server generates task-level user profiles based on the global profile capability results, specifically: The federal coordination server receives business function call requests in real time, and selects all capability tags and values ​​of the target user from the global profile capability results according to the combination rules of the profile task directory to form a capability subset. Priority arbitration is performed on conflicting labels in the capability subset, and task-level user profiles are generated after exponential decay of lagging capability labels.

10. The method for implementing multi-dataset user profiling based on federated learning as described in claim 1, characterized in that: The usage feedback records returned by the received service function form an application result set, specifically: After completing business decisions using task-level user profiles, business functions generate usage feedback records and send these records back to the federated coordination server via a secure channel. The federal coordination server uses the profile task identifier and the target user identifier as keys to associate the usage feedback records with the corresponding task-level user profiles one by one; The associated records, along with the version number of the global profiling capabilities used at the time, are aggregated to form an application result set.

Images